Association of experiencing homelessness with hospital utilization and medical bills among people who have injected drugs: A community-based cohort and data linkage study

Abstract

Background.

Housing instability is common among people who inject drugs and is associated with poor health. We examined the association of homelessness with emergency department (ED) and inpatient (IP) utilization and charges among people who have injected drugs.

Methods.

Participants were enrolled 2015-2024 in ALIVE, a cohort of Baltimore adults who have injected drugs. Past-six-month homelessness was reported at twice-annual study visits. ED and IP admissions and resulting charges to patients were ascertained through linkage to a state hospital billing registry. We estimated differences in rates of admissions and charges per-six-months comparing periods when participants did and did not experience homelessness, and adjusting for confounders such as substance use behaviors using balancing weights.

Results.

Among 1,365 participants’ 9,896 study visits (66% male, 87% age 40-69), mean past-six-month rates of admissions and charges were 0.92 and $1,408 respectively for the ED, and 0.19 and $4,051 for IP. After confounding adjustment, ED admission rates (adjusted rate difference (aRD) = 1.37, 95% CI 0.79 to 2.42), IP admission rates (aRD 0.11, 95% CI 0.04 to 0.19), ED charges (aRD = $1,720, 95% CI $1,027 to $2,726), and IP charges (aRD $2,072, 95% CI $516 to $3,760) were significantly higher during periods of homelessness.

Discussion.

When people who injected drugs were experiencing homelessness, they experienced significantly elevated rates of ED and IP admissions and charges.

Background

People with a history of injecting drugs use hospital-based emergency department (ED) and inpatient (IP) health care at higher rates than people who have not injected drugs.^1-5 Common primary diagnoses at ED admission include drug overdose poisonings, other injuries such as assaults, and psychiatric diagnoses.^3,4 Common diagnoses at IP admission include pneumonias and soft-tissue infections.⁵ In the United States, the ongoing opioid epidemic contributed to rapid growth in drug-related ED and IP admissions over the last decade, particularly during the COVID-19 pandemic.^6,7

Homelessness and housing instability are also common experiences among people who inject drugs. One meta-analysis estimated that the prevalence of homelessness among people who inject drugs was 22% globally and 50% in North America specifically.⁸ A study conducted just before the COVID-19 pandemic of people from Baltimore who have injected drugs found that 7.5% self-reported homelessness in the past-six-months.⁹ Moreover, a recent national study from the National Alliance to End Homelessness found that the number of people in the United States experiencing homelessness has increased sharply since 2018,¹⁰ and that people with disabilities – including substance use disorder –have experienced particularly rapid increases.¹⁰ Poverty, co-occurring mental health issues, a history of incarceration, and poor relations with family all may contribute to risk for homelessness among people who inject drugs.¹¹ In the U.S., these challenges are compounded by the fact that federal law prohibiting discrimination against people with disabilities – such as substance use disorder – explicitly excludes people who continue to use drugs that are illegal, permitting housing providers to discriminate against people who use drugs.¹²

This high prevalence of homelessness in turn contributes to the high burden of disease among people who inject drugs. For example, as described by Arum and colleagues,¹³ among people who inject drugs, homelessness is a risk factor for mortality, health risk behaviors, experiencing barriers to treatment and other essential services, and incident HIV and hepatitis C virus infection.

Of particular concern is the potential effect of homelessness to exacerbate already high rates of ED and IP utilization among people who use drugs. In the general population, homelessness is a well-established risk factor for morbidity and ED use.^20,21 In some settings as many as one in 5 ED admissions are for people experiencing homelessness. ²¹ In fact, a recent meta-analysis of risk factors for incident ED and IP hospital admissions among people who inject drugs found that homelessness was associated with approximately 50% increased risk for an incident ED admission and 60% increased risk for an IP admission.¹

Despite the established link between experiencing homelessness and ED utilization – including among people who inject drugs – important gaps exist in our understanding of this relationship. First, more research has examined ED than IP use. Second, past studies typically focused on risk for a single incident episode of ED and IP care. This approach may miss more dramatic associations among people who use drugs, who have much higher baseline rates of ED and IP utilization than people who do not use drugs. Third, to our knowledge, no studies have examined medical costs associated with ED and IP utilization among people who inject drugs. Understanding these charges is important because most adults with opioid use disorder in the United States are either uninsured or publicly insured.²² Finally, only one study referenced in a recent meta-analysis on homelessness and ED use among people who inject drugs was set in the United States. ¹

In this study, we used data from a cohort study of people who have injected drugs to examine differences in the rate (per-six-months) of hospital ED and IP admissions and charges comparing periods with self-reported homelessness to periods without. We ascertained admissions and associated charges from all Maryland hospitals using linkage to a state registry of hospital bills. This design allows us to address gaps in existing research: we examine both ED and IP admissions using the same methodology rather than ED admissions alone, we examine rates of admission during a six-month period rather than focusing only single admissions, we examine medical charges associated with these admissions in addition to counting them, and we use a U.S.-based cohort.

In addition to overall effects, we examined effects within three sub-groups of people who have injected drugs: those who recently reported any injection drug use, those with recent non-injection drug use only, and those without any recent drug use. This sub-group analysis has important implications for policies that require people who use drugs to stop using or participate in treatment as a condition of receiving housing support. If the association of homelessness with hospital utilization is comparable irrespective of whether or how participants have recently used drugs, this undercuts health-related justifications for making housing contingent on sobriety.

Finally, the COVID-19 pandemic was associated with stark increases in ED and IP admissions for overdoses and other drug-related causes.^6,7 By contrast, the pandemic was also associated with overall decreases in all-cause ED and IP admissions in the full population,^23,24 possibly because patients avoided health care encounters to reduce risk for SARS-CoV-2 infection. Finally, the pandemic was associated with sharp increases in the population prevalence of homelessness.¹⁰ Because of these countervailing trends, we also examined the association of homelessness with each outcome before, during, and after the start of the COVID-19 pandemic, to see if the pre-, inter-, or post-pandemic era modified the association of homelessness with ED and IP admissions among people who have injected drugs. Evidence for effect modification could suggest that, in the wake of COVID-19, the epidemiologic relationship between homelessness and healthcare utilization among people who have used drugs has been altered.

Methods

Participants were enrolled in the AIDS Linked to the Intravenous Experience (ALIVE) study: a community-recruited cohort study of adults who have injected drugs in the Baltimore, Maryland area active since 1988.²⁵ Participants attend twice-annual in-person visits at the ALIVE clinic until death, loss-to-follow-up, or voluntary withdrawal. During each visit, participants complete an interviewer or audio computer assisted interview related to substance use behaviors and provide blood. (See Appendix A: Detailed Methods, A.1. Participants).

To ascertain participants ED and IP utilization, we submitted identifiers for 3,321 ALIVE participants who provided informed consent to “CRISP:” Maryland’s Health Information Exchange (HIE). CRISP used a validated algorithm to link participants to a complete registry of ED and IP admissions and associated charges at non-federal Maryland hospitals. The analytic sample for this analysis consisted of data collected from participants with linked CRISP health care episodes who also had attended at least two ALIVE study visits between April 1, 2015 – September 30, 2024. Visits (n=96) were excluded if, at that study visit, data were missing on any covariates to be included in the analysis. The final sample comprised 1,365 participants who contributed 9,896 study visits. (See Appendix A: Detailed Methods, A.2. Participants).

The primary exposure of interest was self-reported homelessness in the six months preceding the study visit. This study had four primary study outcomes, all ascertained through linkage to CRISP and measured during the six-month period preceding the study visit: (1) The number of hospital ED admissions; (2) The number of hospital IP admissions; (3) The total dollar value of all charges assessed for these ED admissions; and (4) The value of all charges assessed for these IP admissions. In addition to all-cause outcomes, we conducted sub-analyses of admissions and charges with 5 specific drug-related diagnoses present during the episode of care (based on ICD-10 codes): (1) substance use disorders excluding alcohol, (2) alcohol use disorders, (3) psychoses, (4) depression, and (5) drug overdose. The first four were defined using the Elixhauser Comorbidity Index.^26,27 Overdose was defined using the U.S. CDC surveillance definition.²⁸ For effect modification analyses of drug use, we classified visits based on survey responses into three groups: (1) injected drugs; (2) used non-injection drugs only; (3) did not use illicit drugs. For effect modification analyses focused on the COVID-19 pandemic, visits were classified as pre-pandemic (before March 1, 2020); during the pandemic (between March 1, 2020 and May 04, 2023), and post-pandemic (on or after May 05, 2023). Additional covariates included for confounding adjustment are shown in Table 1. (See Appendix A: Detailed Methods, A.3. Measures).

Table 1. Characteristics of participants who have injected drugs (n = 1,365) at bi-annual study visits (n = 9,896) by past-six-homelessness – Baltimore, MD 2015-2024.

Characteristic	Full Sample		No Homelessness		Homelessness
	n	%	n	%	n	%
Route of Administration
Injection	2299	23%	1780	20%	519	55%
Non-Injection Only	2305	23%	2076	23%	229	24%
No Drug Use	5292	53%	5091	57%	201	21%
Pandemic Era
Pre-Pandemic	5344	54%	4841	54%	503	53%
Pandemic	2419	24%	2225	25%	194	20%
Post-Pandemic	2133	22%	1881	21%	252	27%
Heroin Use
No Heroin	6513	66%	6202	69%	311	33%
Heroin	3383	34%	2745	31%	638	67%
Cocaine Use
No Cocaine	6334	64%	6033	67%	301	32%
Cocaine	3562	36%	2914	33%	648	68%
Cigarette Use
No Cigarettes	2598	26%	2474	28%	124	13%
Cigarettes	7298	74%	6473	72%	825	87%
Cannabis Use
No Cannabis	7694	78%	7067	79%	627	66%
Cannabis	2202	22%	1880	21%	322	34%
Alcohol Use
None	5530	56%	5035	56%	495	52%
Moderate	2379	24%	2145	24%	234	25%
Heavy	1987	20%	1767	20%	220	23%
Illicit Painkiller Use
No Painkillers	9363	95%	8513	95%	850	90%
Painkillers	533	5%	434	5%	99	10%
Illicit Tranquilizer Use
No Tranquilizers	9396	95%	8562	96%	834	88%
Tranquilizers	500	5%	385	4%	115	12%
Illicit Stimulant Use
No Stimulants	9770	99%	8853	99%	917	97%
Stimulants	126	1%	94	1%	32	3%
Illicit Sedative Use
No Sedatives	9490	96%	8616	96%	874	92%
Sedatives	406	4%	331	4%	75	8%
Illicit Hallucinogen Use
No Hallucinogens	9808	99%	8886	99%	922	97%
Hallucinogens	88	1%	61	1%	27	3%
HIV Ab Status
HIV−	6927	70%	6191	69%	736	78%
HIV+	2969	30%	2756	31%	213	22%
Employment Status
Not Working	8469	86%	7598	85%	871	92%
Working	1427	14%	1349	15%	78	8%
Self-Reported Health
Excellent/Good/Very Good	6642	67%	6101	68%	541	57%
Fair/Poor	3254	33%	2846	32%	408	43%
Methadone Treatment
No Methadone	6161	62%	5671	63%	490	52%
Methadone	3735	38%	3276	37%	459	48%
Buprenorphine Treatment
No Buprenorphine	8540	86%	7782	87%	758	80%
Buprenorphine	1356	14%	1165	13%	191	20%
Gender
Male	6476	65%	5815	65%	661	70%
Female	3420	35%	3132	35%	288	30%
Age
18-39	446	5%	282	3%	164	17%
40-69	8660	88%	7899	88%	761	80%
70+	790	8%	766	9%	24	3%
Race
Other Race Groups	1480	15%	1051	12%	429	45%
Black	8416	85%	7896	88%	520	55%
Number of ED Visits
0 ED Visits	6001	61%	5618	63%	383	40%
1 ED Visit	2120	21%	1915	21%	205	22%
2+ ED Visits	1775	18%	1414	16%	361	38%
Number of IP Visits
0 IP Visits	8636	87%	7870	88%	766	81%
1 IP Visit	906	9%	784	9%	122	13%
2+ IP Visits	354	4%	293	3%	61	6%

We estimated the “Average Treatment Effect on Treated (ATT) of experiencing homelessness on the past-six-month rate of (1) hospital ED admissions, (2) ED charges, (3) IP admissions, and (4) IP charges. We did so two ways: as the between-group rate difference in average admissions/charges per six-months, and as a rate ratio of average admissions/charges per six-months. We estimated these effects by using entropy-balancing weights to reweight the sample of participant-visits at which participants did not experience homelessness to have the same mean on all potential confounding covariates as the sample of participant visits at which participants did experience homelessness. To quantify uncertainty, we used balanced clustered bootstrapping, with 10,000 bootstrap replicates and clustering at the participant level. We reported the median of these bootstrap replicates as our point estimate, and the 0.025 and 0.975 quantiles as our 95% bootstrap confidence interval.²⁹ To examine effect modification, we stratified by the drug use and pandemic subgroups described above and repeated the analyses within each subgroup. We also repeated the analysis for each contributing cause of interest, counting only outcomes where the relevant contributing causes was indicated by the presence of an ICD-10 code in any position on the bill for that episode. We conducted two sensitivity analyses to explore sensitivity to outliers and our definition of the outcomes: results are described in Appendix B. (See Appendix A: Detailed Methods, A.1. Analytic Approach).

All analyses were conducted in R version 4.4.0, using the “weightit” package to estimate entropy-balancing weights.

Results

Description of Sample

There were 1,365 participants contributing 9,896 study visits. Participants were 66% male, 87% 40-69, and 78% Black. Participants reported past-six-month homelessness at 10% of all study visits. The past-six-month prevalences of drug use by route of administration were 23% injection, 23% non-injection only, and 54% no drug use. 54% of study visits occurred before March 1, 2020, i.e. “pre-pandemic,” 24% between March 1, 2020 and May 04, 2023 i.e. pandemic, and 22% on or after May 05 2023 i.e. post-pandemic. Complete study-visit-level characteristics of the sample are shown in Table 1, and participant-level characteristics in Appendix Table 1.

Mean past-six-month ED admissions and charges were 0.92 and $1,408 respectively. 38% of study visits had an ED admission in the preceding 6 months. Among visits with an ED admission in the past 180 days, the median number of admissions identified during that period was 1 (90^th percentile 4, maximum 126). Further, among visits with an ED admission in the past 180 days, the median amount of associated charges incurred during that period was $1,810 (90^th percentile $8,569, maximum $133,556).

Mean IP admissions and charges were 0.19 and $4,051 respectively. 13% of study visits had an IP admission in the preceding 180 days. Among visits with an IP admission in the past 6 months, the median number of admissions identified during that period was 1 (90^th percentile 3, maximum 11). Further, among visits with an IP admission in the past 180 days, the median amount of associated charges incurred during that period was $20,008 (90^th percentile $70,067, maximum $330,336).

Complete outcomes by homelessness status, route of administration, and pandemic era are shown in Table 2. Histograms for the distribution of each outcome are shown in Appendix Figure 1.

Table 2. Past-six-month emergency department and hospital utilization and charges by housing status, route of drug administration, and pandemic-era among people who have injected drugs.

Subsample	Outcome	Setting	Mean
			All Participants	No Homelessness	Homeless
Full Sample	Number of Admissions	Emergency Department	0.92	0.73	2.66
		Inpatient	0.19	0.17	0.36
	Total Charges	Emergency Department	1408.26	1177.52	3583.62
		Inpatient	4051.12	3764.89	6749.59
Injected Drugs	Number of Admissions	Emergency Department	1.33	0.88	2.86
		Inpatient	0.24	0.20	0.40
	Total Charges	Emergency Department	1940.18	1392.15	3819.75
		Inpatient	4919.29	4182.49	7446.29
Non-Injection Only	Number of Admissions	Emergency Department	1.03	0.82	2.87
		Inpatient	0.18	0.17	0.30
	Total Charges	Emergency Department	1484.75	1252.92	3586.40
		Inpatient	3563.96	3300.05	5956.49
No Drug Use	Number of Admissions	Emergency Department	0.69	0.65	1.90
		Inpatient	0.17	0.17	0.31
	Total Charges	Emergency Department	1143.87	1071.74	2970.76
		Inpatient	3886.14	3808.44	5854.21
Pre-Pandemic	Number of Admissions	Emergency Department	1.02	0.85	2.69
		Inpatient	0.21	0.19	0.39
	Total Charges	Emergency Department	1450.98	1206.89	3800.25
		Inpatient	3879.02	3647.38	6108.39
Pandemic	Number of Admissions	Emergency Department	0.76	0.60	2.57
		Inpatient	0.17	0.16	0.32
	Total Charges	Emergency Department	1362.47	1163.57	3643.59
		Inpatient	4209.31	3813.76	8745.83
Post-Pandemic
	Number of Admissions	Emergency Department	0.85	0.61	2.67
		Inpatient	0.18	0.16	0.31
	Total Charges	Emergency Department	1353.17	1118.46	3105.06
		Inpatient	4302.88	4009.52	6492.64

Association of Homelessness with Hospital Utilization

As compared to study visits when participants did not report homelessness, visits when participants reported past-six-month homelessness had average higher rates per-six-months of ED admissions (0.73 vs 2.66), IP admissions (0.17 vs 0.36), ED charges ($1,177 vs $3,583), and IP charges ($3,765 vs $6,750). In adjusted analysis, all outcomes were significantly higher during periods of homelessness. Specifically, we estimated that past-six-month homelessness was associated with an average increase of 1.37 ED admissions (95% bootstrap confidence interval [BI] 0.79 to 2.42), 0.11 IP admissions (95% BI 0.04 to 0.19), $1,720 in ED charges (95% BI $1,027 - $2,726), and $2,072 in IP charges (95% BI $516 - $3,760). Figure 1 shows the full bootstrap distribution for the estimated mean of each outcome in each group. Complete unadjusted and adjusted rate differences are shown in Table 3. Complete rate ratios are shown in Appendix Table 2.

Figure 1. Clustered bootstrap distribution (10,000 replicates) of estimates of emergency department and inpatient hospital utilization by past-six-month homelessness among people who have injected drugs.

Table 3. Association of past-six-month homelessness with past-six-month emergency department and inpatient hospital utilization, by route of drug administration and pandemic era -- crude and adjusted difference in admissions per 6 months.

Subsample	Outcome	Setting	Rate Difference (95% Bootstrap Interval)
			Crude	Adjusted
Full Sample	Number of Admissions	Emergency Department	1.87 (1.17 - 3.16)	1.37 (0.79 - 2.42)
		Inpatient	0.18 (0.11 - 0.26)	0.11 (0.04 - 0.19)
	Total Charges	Emergency Department	$2,342 ($1,558 - $3,608)	$1,720 ($1,027 - $2,726)
		Inpatient	$2,959 ($1,493 - $4,630)	$2,072 ($516 - $3,760)
Injection	Number of Admissions	Emergency Department	1.91 (1.07 - 3.45)	1.52 (0.74 - 2.97)
		Inpatient	0.2 (0.1 - 0.31)	0.14 (0.02 - 0.26)
	Total Charges	Emergency Department	$2,351 ($1,338 - $3,955)	$1,906 ($916 - $3,420)
		Inpatient	$3,227 ($1,054 - $5,649)	$2,830 ($554 - $5,382)
Non-Injection Only	Number of Admissions	Emergency Department	1.95 (0.9 - 3.95)	1.46 (0.54 - 3.12)
		Inpatient	0.13 (0.03 - 0.24)	0.06 (−0.06 - 0.17)
	Total Charges	Emergency Department	$2,266 ($1,149 - $3,926)	$1,616 ($590 - $3,042)
		Inpatient	$2,608 ($470 - $5,114)	$1,226 (−$1,446 - $3,950)
No Drug Use	Number of Admissions	Emergency Department	1.24 (0.72 - 1.87)	0.99 (0.51 - 1.61)
		Inpatient	0.14 (0.02 - 0.29)	0.09 (−0.01 - 0.22)
	Total Charges	Emergency Department	$1,872 ($1,026 - $2,895)	$1,497 ($674 - $2,511)
		Inpatient	$1,933 (−$759 - $5,357)	$946 (−$1,775 - $4,256)
Pre-Pandemic	Number of Admissions	Emergency Department	1.84 (1.37 - 2.34)	1.31 (0.9 - 1.75)
		Inpatient	0.21 (0.1 - 0.32)	0.12 (0.02 - 0.23)
	Total Charges	Emergency Department	$2,583 ($1,846 - $3,386)	$1,850 ($1,151 - $2,604)
		Inpatient	$2,420 ($599 - $4,469)	$1,945 ($182 - $3,961)
Pandemic	Number of Admissions	Emergency Department	1.89 (0.72 - 4.24)	1.54 (0.38 - 3.84)
		Inpatient	0.16 (0.03 - 0.32)	0.07 (−0.08 - 0.23)
	Total Charges	Emergency Department	$2,347 ($875 - $5,090)	$1,811 ($338 - $4,480)
		Inpatient	$4,810 ($1,219 - $9,476)	$2,606 (−$1,599 - $7,421)
Post-Pandemic	Number of Admissions	Emergency Department	1.97 (0.68 - 4.56)	1.57 (0.42 - 3.84)
		Inpatient	0.15 (0.05 - 0.26)	0.11 (0 - 0.23)
	Total Charges	Emergency Department	$1,865 ($692 - $3,940)	$1,470 ($348 - $3,246)
		Inpatient	$2,432 ($22 - $5,163)	$1,765 (−$965 - $4,635)

Effects statistically significant at the p<0.05 level are bolded.

Stratification by Route of Drug Administration

Outcome rates corresponding to time periods with injection drug use, non-injection drug use, and no illicit drug use were, respectively: ED admissions (1.33 vs 1.03 vs 0.69); IP admissions (0.24, vs 0.18, vs 0.17); ED charges ($1,940, vs $1,485 vs $1,144); and IP charges ($4,919 vs $3,564 vs $3,886). The estimated association of homelessness with each outcome was qualitatively similar across all three groups (Table 3, Appendix Table 2), and in all cases we found no evidence for statistically significant interaction. In other words, we found no evidence that route of drug administration modified the association of homelessness any hospital utilization outcome (Appendix Table 3).

Stratification by Pandemic Era

Outcome rates for study visits before the pandemic as compared to during and after the pandemic were, respectively: ED admissions (1.02 vs 0.76 vs 0.85); IP admissions (0.21 vs. 0.17 vs. 0.18); ED charges ($1,451 vs $1,362 vs $1,353); and IP charges ($3,879 vs $4,209 vs $4,303). The estimated association of homelessness with each outcome was qualitatively similar across both groups (Table 3, Appendix Table 2), and in all cases we found no evidence for statistically significant interaction. In other words, we found no evidence that whether study visits occurred before or after the start of the COVID-19 pandemic modified the association of homelessness any hospital utilization outcome (Appendix Tables 4-6).

Contributing-Cause-Specific Results

ED admissions with ICD-10 diagnosis codes for substance use disorders excluding alcohol, alcohol use disorders, psychoses, depression, and drug overdose comprised 26%, 6%, 6%, 9%, and 4% of all ED admissions respectively. IP admissions with ICD-10 diagnosis codes for substance use disorders excluding alcohol, alcohol use disorders, psychoses, depression, and drug overdose comprised 70%, 19%, 7%, 29%, and 5% of all IP admissions respectively. In adjusted analysis, past-six-month homelessness was associated with significantly higher rates of ED admissions for all five causes examined, but with higher rates of IP admissions for only drug-related admissions and psychoses. Detailed estimates are shown in Figure 2 and presented in Appendix Table 7.

Figure 2. Adjusted relative rate and 95% bootstrap confidence interval for association of homelessness with emergency department and inpatient hospital admissions and charges with specific conditions present on discharge.

Sensitivity Analyses

Our ED-related findings were also robust in all sensitivity analyses. Our IP results were attenuated in our second sensitivity analysis, where the exposure was lagged with respect to the outcome instead of using overlapping time periods. For details, see Appendix B. Sensitivity Analysis Results, and Appendix Tables 8-10.

Discussion

In this cohort of people who have injected drugs, after adjusting for a wide range of potential confounders, we observed that participants experiencing homelessness during a six-month period had nearly double the rate of ED admissions and charges during that period, and a 50% increase in IP admissions and charges, as compared to periods when participants did not experience homelessness. This was equivalent to an excess of 1.37 ED admissions on average resulting in $1,720 more in charges on average, as well as an excess of 0.11 IP admissions on average resulting in $2,072 more in charges on average. With a 10% visit-level prevalence of past-six-month homelessness, our results imply that experiencing homelessness is a critical contributor to high levels of hospital utilization among people who have used drugs. In general, we cannot infer causation from this observational study. However, our findings are consistent with the hypothesis that experiencing homelessness itself – rather than simply the fact that people who are homeless are more likely to be actively using drugs or experiencing worse underlying health – contributes to increased risk for emergency care and, to a lesser extent and with somewhat weaker evidence, to inpatient hospital utilization among people who have injected drugs.

Moreover, because participants’ illicit drug use behaviors varied over time, we were able to conduct analyses stratified by participants’ past-six-month drug use. Overall risk for ED and IP utilization and ED chargers were all lower during periods when participants exclusively used non-injection drugs as compared to injection drugs, and in turn unsurprisingly lower during periods with no drug use. (IP charges followed a different trend.) However, the adjusted relative association of homelessness with all four hospital utilization outcomes was similar across all three drug use strata, and we found no evidence that route of administration (comparing injection and non-injection) modified the association of homelessness with hospital utilization or charges. This is important, because it implies that, not only is homelessness associated with increased ED and IP utilization and charges for people who have injected drugs, but these associations remain present and are comparable irrespective of whether or how participants have recently used drugs.

This last finding is consistent with experimental studies of housing interventions for people experiencing homelessness with behavioral health conditions. Specifically, meta analyses of various experimental interventions to provide housing to people with behavioral health conditions have found that acquiring housing (a) increased participants’ days housed, ^30,31 (b) reduced ED and IP utilization,³¹ but (c) did not reduce substance use.^30,31 The findings of our cohort study are consistent with these experimental studies because we show that, when people who have injected drugs acquire housing, they tend to have lower rates of acute care utilization, and this is true even among people who continue to use injection and non-injection drugs while housed. Together, these findings suggest that the benefits of housing are partly independent of drug use behaviors.

Limitations

This study has a number of limitations. First, while participants are scheduled to attend ALIVE study visits every six months, they may experience longer gaps or miss study visits entirely. In this analysis, median times since last visit for visits with and without homelessness reported were 186 and 188 days respectively, but 90^th percentile times were 406 and 526 days respectively. This suggests that participants experiencing homelessness may have been more likely to miss study visits, which could induce selection bias. This also means that, for some participants, their “prior ED and IP use” covariates are drawn from more distant time periods than others. Second, we limited our sample to participants who had at least one episode of hospital-based health care identified by CRISP in the 9 year period Oct 1 2015 through September 30, 2024, or at least one episode of health care at the broader set of any CRISP-participating health care facility in Maryland during the 5 year period Jan 1, 2019 – December 31, 2024. If some participants truly received no health care during the entire time period for which we conducted CRISP linkage, our estimates of overall ED and IP utilization would be upwardly biased. Third, we also cannot independently verify the accuracy of CRISP’s matching algorithm, although the algorithm is sufficiently accurate to be used as the foundation of Maryland’s HIE inter-hospital patient record sharing, and our manual verification of a small sample of ED and IP admissions confirmed those admissions did occur. CRISP also has a known limitation of providing data only on admissions to hospitals located in Maryland. Fourth, our exposure measurements may be subject to imprecise measurement if participants were homeless but only briefly during a six-month period. Fifth, our study is not a randomized trial and may be biased by residual confounding from unmeasured variables. Sixth, the “charges” outcome we analyzed in this analysis is limited because the amount listed as the charge by a hospital is different than the amount actually paid by each of three possible payers: the patient, the patient’s insurance (which may be public or private), or the hospital system itself (which may absorb the uncompensated care as a loss). We do not know which entity ultimately bore the costs of the assessed charges. Future research on the financial costs of homelessness-associated hospital use can inform cost-benefit analyses examining housing and other public health investments. Seventh, our IP results were not robust to every sensitivity analysis. A greater body of literature supports the association of homelessness with ED utilization as compared to IP utilization. We hope future research will continue to provide more precise estimates of the association with homelessness with IP utilization. Eighth, as shown in Appendix Figure 1, the distribution of outcomes in this analysis is extremely skewed. The use of bootstrapping – and the use of the median of the bootstrap distribution as a point estimate – is intended to avoid parametric assumptions about these skewed outcome distributions and reduce the influence of extreme values on our conclusions. It is important to consider the entire range of values within the confidence interval as a plausible estimate of the effects of homelessness. Ninth, because this was an exploratory study, we did not adjust for multiple comparisons across the analyses we ran. Finally, findings may not generalize to other geographies, to younger populations, or to people who have used drugs but never injected.

Conclusion

Among this sample of people who have injected drugs, we found that experiencing homelessness was associated with increased ED and IP admissions resulting in thousands of dollars in additional charges over a six-month period. These associations persisted after adjusting for participants’ substance use, recent health care utilization, and other potential confounders, and were largely consistent irrespective of whether or how participants recently used illicit drugs. Together, these findings are consistent with the hypothesis that homelessness is an independent contributor to ED and hospital utilization and associated medical bills among people who currently or formerly inject drugs.

Our findings underscore the public health and potential cost-averting benefits of housing for people who have injected drugs. As noted above, experimental studies have shown that “housing first” interventions – which provide housing to adults with behavioral health problems without first requiring cessation of drug use or conformity to other eligibility criteria – reduce ED and hospital use among recipients even though these interventions may not reduce drug use.^30,31 Our findings build on this research to show that, even among people who continue to inject drugs – a cohort with a high burden of disease and behavioral risks, and very high rates of hospital utilization – the public health and health care cost-averting benefits of stable housing persist, and are comparable to benefits observed among people who are no longer injecting drugs. We believe these findings have implications for two areas of policy. The first are “drug-free housing” policies which, though sometimes intended to promote a healthy living environment, may exacerbate homelessness among people who use drugs. Our findings illustrate how such policies, should they increase homelessness among people who inject drugs, may be associated with inadvertent public health harms, with costs ultimately borne by the broader public. The second implication pertains to public investments in permanent supportive housing. This is because our findings imply that, if effective, public investments in housing for people who have injected drugs may in turn generate savings to the public in the form of reduced health care spending, but these savings are not necessarily obvious if only housing expenditures are considered. Our unique linked analysis makes these potential savings more transparent.

Appendix A. Detailed Methods

A.1. Participants

Participants were enrolled in the AIDS Linked to the Intravenous Experience (ALIVE) study: a community-recruited cohort study of adults who have injected drugs in the Baltimore, Maryland area. A full description of the ALIVE study methods are available.²⁵ Briefly, ALIVE first recruited participants between 1988-1989, and conducted six subsequent rounds of recruitment (1994-1995, 1998, 2000, 2005-2008, 2015-2018, 2023-ongoing). ALIVE participants were recruited from within the community at sites and locations frequented by people who use drugs such as the syringe service program, drop-in centers, service centers, laundromats, or grocery stores. They were followed twice annually through in-person visits at the ALIVE clinic until death, loss-to-follow-up, or voluntary withdrawal from the study. During each visit, participants complete an interviewer or audio computer assisted interview related to past-six-month drug consumption behaviors, treatment for substance use disorder, medical information and healthcare utilization, health risk behaviors, and demographics. In addition to the questionnaires, participants provide a blood sample to assess HIV and hepatitis C virus (HCV) status.

A.2. Study Data

ALIVE began requesting participant consent for linkage to external registries in 2005. For this study, we submitted identifiers for all 3,321 ALIVE participants who have consented to registry linkage to the Chesapeake Regional Information System for Our Patients (CRISP). CRISP is Maryland’s Health Information Exchange (HIE). Designed to facilitate linkage of medical records between health systems, CRISP maintains a registry of all patients who receive care at all non-federal hospitals in Maryland, as well as a subset of outpatient practices that choose to participate. CRISP linked ALIVE participants to this patient registry using its internal, validated, HIPAA-compliant probabilistic algorithm. Using this linkage, CRISP provided participants’ complete healthcare utilization records from two data systems: (1) the “encounter notification” system used for tracking health care encounters at all CRISP-participating health care practices, for the period Jan 1, 2019 through Dec 31 2024; and (2) administrative discharge and billing data submitted by all Maryland acute-care hospitals to the State’s Health Services Cost Review Commission (HSCRC), for the period Oct 1 2015 through September 30, 2024, commonly referred to as HSCRC’s “case mix” data.³²

Of 3,321 participants submitted to CRISP, 346 were unable to be linked by CRISP to any person in its patient registry. Of these 346, 18 had died prior to 2015; an additional 15 were living but had not attended an ALIVE study visit since before 2015, and thus may have moved out of Maryland. An additional 702 participants were linked to CRISP’s patient registry but did not have a single episode of health care recorded in any of the datasets provided by CRISP. Of these, 411 had died prior to 2015; an additional 195 were living but had not attended an ALIVE visit since 2015, and thus may have moved out of Maryland. We then randomly selected three participants who were linked in CRISP, had no episodes of health care returned by CRISP, but had attended at least one ALIVE study visit since 2015. For these three participants, we conducted manual medical record review through Johns Hopkins electronic medical records system, and for all three identified health care episodes that should have been captured by our CRISP linkage but were not. Consequently, we decided to assume that all of these individuals who were linked by CRISP but did not have any returned health care episodes represented linkage errors rather than “true zeros,” and excluded them. Finally, there were 9 participants that CRISP returned care episodes for, but at least one of those care episodes occurred after the participant’s date of death (as provided by NDI); we assumed these represented linkage errors, and excluded these participants as well.

Consequently, in this analysis, we began with a sample of 2,264 ALIVE participants who were (1) living on Oct 1, 2015 and (2) had at least one linked health care episode in any dataset returned by CRISP. Other than for determining inclusion criteria, only the “case mix” dataset was used for this analysis, because only that case mix dataset contained all ED and IP hospital billing records – the primary outcomes of this study – which contained the complete ICD-10 diagnosis and billing information required for constructing the study outcomes.

The analytic sample for this analysis consisted of data collected from participants with linked CRISP health care episodes who also had attended at least two ALIVE study visits (to facilitate “lagged” covariates, see “Measures”) between April 1, 2015 – September 30, 2024. Visits (n=96) were excluded if, at that study visit, data were missing on any covariates to be included in the analysis. The final sample comprised 1,365 participants who contributed 9,896 study visits.

A.3. Measures

A.3.1. Homelessness.

The primary exposure of interest was self-reported homelessness in the six months preceding the study visit. Participants were asked, “In the past six months have you…” (1) “been homeless?” or (2) “…slept in a shelter, park, abandoned building, in the street, or in a bus or train station for at least one night because you had nowhere else to stay?” In a 2023 questionnaire revision, the second question was modified to add “in a tent” as an option. Participants who endorsed either option were classified as having experienced homelessness in the past six months.

A.3.2. Hospital Use and Charges.

All outcome measures were derived from the “case mix” data provided by CRISP. This study had four primary study outcomes, all measured during the six-month period preceding the study visit: (1) The number of hospital ED admissions; (2) The number of hospital IP admissions; (3) The total dollar value of all charges assessed for these ED admissions; and (4) The total dollar value of all charges assessed for these IP admissions.

In addition to all-cause outcomes, we conducted sub-analyses of admissions and charges with specific diagnoses present during the episode of care, as indicated by the presence of an ICD-10 code in any position on the bill for that episode. We examined five behavioral health conditions and their sequelae that are closely related to injection drug use: (1) substance use disorders excluding alcohol, (2) alcohol use disorders, (3) psychoses, (4) depression, and (5) drug overdose. The first four were defined using the set of ICD-10 codes used to identify those conditions in the Elixhauser Comorbidity Index.^26,27 Overdose was defined using the set of ICD-10 codes recommended by the U.S. CDC for overdose surveillance.²⁸ Note that these outcomes are not mutually exclusive, because each admission can have multiple associated ICD-10 codes.

A.3.3. Route of Drug Administration.

Participants were asked “Thinking back over the last six months, and all the different drugs you may have injected, what is the total number of times that you injected drugs?” Participants were also asked “Thinking back over the last six months, and all the different drugs you may have taken by swallowing, snorting, sniffing or smoking, (other than injecting), what is the total number of times that you took drugs?” Based on their responses, study visits were classified into one of three mutually exclusive groups: (1) injected drugs; (2) used non-injection drugs only; (3) did not use illicit drugs.

A.3.4. COVID-19 Pandemic.

Visits were classified as having occurred before the COVID-19 pandemic if they occurred between April 1, 2015 and March 1, 2020; during the COVID-19 pandemic if they occurred between March 1, 2020 and May 04, 2023 (the date the World Health Organization [WHO] declared an end to the COVID-19 global health emergency); and after the COVID-19 pandemic if they occurred on or after May 05, 2023 to September 30, 2024 (the administrative censoring date for our study).

A.3.5. Other covariates.

We included the following additional covariates for confounding adjustment, all based on self-report, ascertained at the same visit at which homelessness status was assessed, and covering the same six-month period as the homelessness question except as otherwise noted: heroin use (yes/no), cocaine use (yes/no), cigarette use (yes/no), marijuana use (yes/no), alcohol use severity based on the Alcohol Use Disorders Identification Test (AUDIT) (none/moderate/heavy),³³ non-prescribed pain-reliever use (yes/no); non-prescribed stimulant use (yes/no); non-prescribed sedative use (yes/no); non-prescribed tranquilizer use (yes/no); non-prescribed hallucinogen use (yes/no); HIV antibody status, based on a blood specimen collected at the study visit, and as assessed by an enzyme-linked immunosorbent assay and confirmed by Western blot (yes/no); HCV antibody status, based on a blood specimen collected at the study visit (yes/no); employment status (not working/working); self-reported general health status (Excellent, Very Good, Good/Fair, poor); self-reported pain (None, Very Mild, Moderate/Moderate, Severe, Very Severe); received methadone treatment (yes/no); received buprenorphine treatment (yes/no); gender (male/female); age group (18-39/40-69/70+); race (all races other than Black/Black); number of ED admissions in the six months preceding the immediate prior study visit, i.e. the ED outcome lagged by one visit and categorized (0, 1, 2+); number of IP admissions in the six months preceding the immediate prior study visit, i.e. the IP outcome lagged by one visit and categorized (0, 1, 2+).

A.4. Analytic Approach

A.4.1. Main analysis.

The goal of the analysis was to estimate, among people who have injected drugs, the effect of experiencing homelessness on the past-six-month rate of (1) hospital ED admissions, (2) ED charges, (3) IP admissions, and (4) IP charges, comparing the observed rate of each outcome among homeless individuals to a counterfactual scenario in which those individuals had not experienced homelessness. In technical terms, we estimated an “Average Treatment Effect on Treated” [ATT]. This means our analysis tries to answer the question: “How would the hospital utilization of people who inject drugs who experienced homelessness differ, on average, had these people, counter to fact, not experienced homelessness?”³⁴ The ATT is relevant when examining the effect of an exposure that would typically be experienced by individuals demographically or behaviorally like the individuals who actually experienced that exposure. ³⁴

We estimate the ATTs two ways: as the between-group rate difference in average admissions/charges per six-months, and as a rate ratio of average admissions/charges per six-months. We estimated these effects by using entropy-balancing weights to reweight the sample of participant-visits at which participants did not experience homelessness to have the same mean (within 0.001 percentage points) on all potential confounding covariates as the sample of participant visits at which participants did experience homelessness.³⁵ Potential confounders included illicit drug use status by route of administration, pandemic era, and all covariates listed above under “other covariates” above. We then computed the mean of each outcome in the homeless group and the re-weighted not homeless group, and calculated the ATT as a rate difference and rate ratio between groups simply by taking the between-group difference and ratio in weighted means respectively. We chose this weighting approach – which does not rely on the parametric assumptions of regression modeling – because, as shown in Appendix Figure 1, the distribution of all outcomes was extremely right-skewed.^36,37 Consequently, we preferred an analytic approach that made few or no parametric assumptions about the distribution of the outcome. An additional benefit of our weighting approach is that it is then simple to present both rate differences and rate ratios as measures of the ATT. The former helps estimate the absolute size of the effect, with more straightforward interpretation for public health impact. However, because the latter essentially scales each effect to be relative to the rate in the control group, it is useful for comparing effect sizes across outcomes which have different base rates. This is particularly useful for the analysis of cause-specific admissions, which have dramatically different base rates, and hence we present only rate ratios. For all effects, we report both crude and adjusted (i.e. weighted) estimates.

To quantify uncertainty, we used balanced clustered bootstrapping, with 10,000 bootstrap replicates and clustering at the participant level. We report the median of these bootstrap replicates as our point estimate (as in “bagging”³⁸), and the 0.025 and 0.975 quantiles as our 95% bootstrap confidence interval.²⁹

A.4.2. Subgroup analyses.

To examine effect modification by route of drug administration, we stratified by past-six-month route of administration and repeated the analysis – including confounding adjustment for the same set of confounders using balancing weights and bootstrapping to quantify uncertainty – within each sub-group.³⁹ (For the sub-group analysis of visits when participants reported neither injection nor non-injection drug use, we did not adjust for heroin use, cocaine use, or non-prescribed use of pain-relievers, stimulants, sedatives, tranquilizers, or hallucinogens, because all were not present by definition.) Similarly, to examine effect modification by pandemic era, we stratified on pandemic era, and repeated the analysis within each subgroup. The result of this analysis is a bootstrap distribution of estimates for the association of homelessness with each outcome within each sub-group. As with our main analysis, we report effect estimates on both the rate-difference and rate ratio scale for each strata-specific estimate.

In order to test for effect modification by subgroup, we then computed the difference in strata-specific effect estimates between strata. (For rate ratio estimates, we took the logarithm of these ratios before computing this difference, hence we report difference in log rate ratio.) This process was repeated in each bootstrapped sample, to produce a full bootstrap distribution for the difference in effect sizes between strata. We determined statistically significant evidence of interaction to be cases where the 95% bootstrap interval for this difference in effects between strata did not contain 0.

(Note that we did not include the sample of visits with no injection in our formal interaction tests because, as noted, we adjusted for a smaller set of confounders in this subsample than in the samples that used injection and exclusive non-injection drugs respectively.)

A.4.3. Contributing-cause-specific admissions.

We then repeated the analysis described above for each contributing cause of interest, counting only outcomes where the relevant contributing causes was indicated by the presence of an ICD-10 code in any position on the bill for that episode. Note that, because rates of admissions and charges vary substantially between contributing causes, effect estimates are presented only on the ratio scale, to facilitate direct comparisons of effect sizes between causes.

A.4.4. Sensitivity analyses.

We conducted two sensitivity analyses.

First, as noted above, the distribution of all outcomes was highly right-skewed, so we examined the sensitivity of our results to a small number of six-month-periods when participants had an extremely high number of admissions. We conducted a sensitivity analysis where we excluded participant visits where the past-six-month period had more than 20 ED admissions or more than 5 IP admissions.

Second, our definition of the outcome was ED and IP admissions in the six months preceding the study visit. We chose this approach so that the period when participants reported homelessness (past-six-months) would overlap with the period when we measured the outcome, since the literature (described above) suggests that it is during the periods when participants are homeless that their ED utilization will be elevated, not after. However, our approach also creates a limitation, which is that we cannot establish temporality, whereby onset of homelessness precedes the outcome. For this reason, we recalculated the outcome using the number of ED and IP admissions and charges in the six months following the participants’ study visit. We then re-ran the main analysis using this alternative definition of the outcome.

All analyses were conducted in R version 4.4.0, using the “weightit” package to estimate entropy-balancing weights.³⁷

Appendix B. Sensitivity Analysis Results

Sensitivity Analysis

In our first sensitivity analysis, we excluded study 43 visits where the past-six-month period had more than 20 ED admissions or more than 5 IP admissions. Results on the difference scale were attenuated as compared to the primary analysis for all outcomes, reflecting the lower outcome rate in this sensitivity analysis that resulted from excluding outliers. However, results on the ratio scale were qualitatively similar in magnitude in the sensitivity analysis to the primary analysis (Appendix Table 8).

In our second sensitivity analysis, we redefined our outcome to be the number of ED or IP visits or charges in the six months following each study visit (rather than the 6 months preceding each visit.) In adjusted analysis, the estimated associations of homelessness with ED utilization and costs were comparable to adjusted estimates from our main analysis, and remained statistically significant. By contrast, adjusted associations with IP admissions were, as compared to our main analysis, attenuated and no longer significant, though in the same direction (Appendix Tables 9 and 10).

Appendix Table 1. Characteristics of participants who have injected drugs (n = 1,365) at first study visits.

Characteristic	Full Sample		No Homelessness		Homelessness
	n	%	n	%	n	%
Route of Administration
Injection	512	38%	366	32%	146	68%
Non-Injection Only	302	22%	267	23%	35	16%
No Drug Use	551	40%	518	45%	33	15%
Pandemic Era
Pre-Pandemic	1162	85%	1016	88%	146	68%
Pandemic	25	2%	20	2%	5	2%
Post-Pandemic	178	13%	115	10%	63	29%
Heroin Use
No Heroin	733	54%	677	59%	56	26%
Heroin	632	46%	474	41%	158	74%
Cocaine Use
No Cocaine	742	54%	686	60%	56	26%
Cocaine	623	46%	465	40%	158	74%
Cigarette Use
No Cigarettes	290	21%	260	23%	30	14%
Cigarettes	1075	79%	891	77%	184	86%
Cannabis Use
No Cannabis	1038	76%	908	79%	130	61%
Cannabis	327	24%	243	21%	84	39%
Alcohol Use
None	698	51%	593	52%	105	49%
Moderate	352	26%	292	25%	60	28%
Heavy	315	23%	266	23%	49	23%
Illicit Painkiller Use
No Painkillers	1230	90%	1043	91%	187	87%
Painkillers	135	10%	108	9%	27	13%
Illicit Tranquilizer Use
No Tranquilizers	1220	89%	1041	90%	179	84%
Tranquilizers	145	11%	110	10%	35	16%
Illicit Stimulant Use
No Stimulants	1326	97%	1122	97%	204	95%
Stimulants	39	3%	29	3%	10	5%
Illicit Sedative Use
No Sedatives	1253	92%	1066	93%	187	87%
Sedatives	112	8%	85	7%	27	13%
Illicit Hallucinogen Use
No Hallucinogens	1341	98%	1133	98%	208	97%
Hallucinogens	24	2%	18	2%	6	3%
HIV Ab Status
HIV−	1003	73%	814	71%	189	88%
HIV+	362	27%	337	29%	25	12%
Employment Status
Not Working	1163	85%	969	84%	194	91%
Working	202	15%	182	16%	20	9%
Self-Reported Health
Excellent/Good/Very Good	890	65%	764	66%	126	59%
Fair/Poor	475	35%	387	34%	88	41%
Methadone Treatment
No Methadone	806	59%	696	60%	110	51%
Methadone	559	41%	455	40%	104	49%
Buprenorphine Treatment
No Buprenorphine	1132	83%	967	84%	165	77%
Buprenorphine	233	17%	184	16%	49	23%
Gender
Male	900	66%	736	64%	164	77%
Female	465	34%	415	36%	50	23%
Age
18-39	138	10%	80	7%	58	27%
40-69	1190	87%	1035	90%	155	72%
70+	37	3%	36	3%	1	0%
Race
Other Race Groups	307	22%	199	17%	108	50%
Black	1058	78%	952	83%	106	50%
Number of ED Visits
0 ED Visits	769	56%	686	60%	83	39%
1 ED Visit	288	21%	244	21%	44	21%
2+ ED Visits	308	23%	221	19%	87	41%
Number of IP Visits
0 IP Visits	1168	86%	996	87%	172	80%
1 IP Visit	141	10%	113	10%	28	13%
2+ IP Visits	56	4%	42	4%	14	7%

Appendix Table 2. Association of past-six-month homelessness with past-six-month emergency department and inpatient hospital utilization, by route of drug administration and pandemic era -- crude and adjusted ratio of admissions per 6 months.

Subsample	Outcome	Setting	Rate Ratio (95% Bootstrap Interval)
			Crude	Adjusted
Full Sample	Number of Admissions	Emergency Department	3.54 (2.61 - 5.16)	2.12 (1.67 - 2.71)
		Inpatient	2.06 (1.63 - 2.57)	1.44 (1.13 - 1.85)
	Total Charges	Emergency Department	3 (2.31 - 4.02)	1.96 (1.57 - 2.43)
		Inpatient	1.79 (1.37 - 2.3)	1.45 (1.1 - 1.9)
	Number of Admissions
Injection		Emergency Department	3.15 (2.2 - 4.78)	2.21 (1.57 - 3.26)
	Total Charges	Inpatient	2.01 (1.45 - 2.74)	1.53 (1.05 - 2.27)
		Emergency Department	2.7 (1.93 - 3.84)	2.05 (1.46 - 2.89)
	Number of Admissions	Inpatient	1.78 (1.23 - 2.53)	1.62 (1.1 - 2.45)
Non-Injection Only	Total Charges	Emergency Department	3.39 (2.12 - 5.24)	2.13 (1.43 - 3.25)
		Inpatient	1.77 (1.19 - 2.57)	1.23 (0.81 - 1.83)
	Number of Admissions	Emergency Department	2.83 (1.91 - 4.04)	1.87 (1.29 - 2.67)
		Inpatient	1.8 (1.13 - 2.73)	1.26 (0.75 - 2.07)
	Total Charges
No Drug Use		Emergency Department	2.92 (2.08 - 3.98)	2.12 (1.56 - 2.92)
	Number of Admissions	Inpatient	1.84 (1.09 - 2.8)	1.44 (0.93 - 2.02)
		Emergency Department	2.75 (1.92 - 3.81)	2.05 (1.43 - 2.88)
	Total Charges	Inpatient	1.51 (0.81 - 2.48)	1.2 (0.65 - 1.97)
Pre-Pandemic	Number of Admissions	Emergency Department	3.18 (2.57 - 3.86)	1.96 (1.61 - 2.36)
		Inpatient	2.11 (1.54 - 2.8)	1.44 (1.05 - 1.98)
	Total Charges	Emergency Department	3.14 (2.48 - 3.9)	1.96 (1.54 - 2.48)
		Inpatient	1.67 (1.16 - 2.3)	1.47 (1.04 - 2.08)
	Number of Admissions
Pandemic		Emergency Department	4.13 (2.19 - 8.09)	2.6 (1.37 - 5.07)
	Total Charges	Inpatient	2.01 (1.2 - 3.19)	1.27 (0.73 - 2.11)
		Emergency Department	3.01 (1.72 - 5.43)	2.07 (1.18 - 3.75)
	Number of Admissions	Inpatient	2.27 (1.3 - 3.67)	1.44 (0.78 - 2.5)
Post-Pandemic	Total Charges	Emergency Department	4.24 (2.22 - 7.16)	2.54 (1.47 - 4.37)
		Inpatient	1.94 (1.31 - 2.8)	1.57 (1 - 2.54)
	Number of Admissions	Emergency Department	2.69 (1.63 - 4.3)	1.98 (1.22 - 3.18)
		Inpatient	1.61 (1.01 - 2.44)	1.38 (0.83 - 2.3)

Appendix Table 3. Estimated difference in association of homelessness with hospital utilization and charges between study visits with injection drug use and participants with exclusive non-injection drug use.

Outcome	Setting	Measure of Association	Interaction Estimate
Number of Admissions	Emergency Department	Rate Difference	0.07 (−1.79 - 1.84)
		Logarithm of Rate Ratio	0.04 (−0.52 - 0.6)
	Inpatient	Rate Difference	0.08 (−0.08 - 0.25)
		Logarithm of Rate Ratio	0.22 (−0.32 - 0.8)
Total Charges	Emergency Department	Rate Difference	288.92 (−1460.97 - 2103.72)
		Logarithm of Rate Ratio	0.1 (−0.4 - 0.6)
	Inpatient	Rate Difference	1577.32 (−1871.54 - 5297.09)
		Logarithm of Rate Ratio	0.25 (−0.37 - 0.92)

Appendix Table 4. Estimated difference in association of homelessness with hospital utilization and charges between study visits before the COVID-19 pandemic and during COVID-19 pandemic (March 1, 2020).

Setting	Outcome	Measure of Association	Interaction Estimate
Number of Admissions	Emergency Department	Rate Difference	0.2 (−1.08 - 2.59)
		Logarithm of Rate Ratio	0.27 (−0.39 - 0.99)
	Inpatient	Rate Difference	−0.05 (−0.24 - 0.15)
		Logarithm of Rate Ratio	−0.13 (−0.77 - 0.48)
Total Charges	Emergency Department	Rate Difference	−43.48 (−1759.32 - 2742.33)
		Logarithm of Rate Ratio	0.05 (−0.56 - 0.69)
	Inpatient	Rate Difference	678.79 (−4002.22 - 5813.68)
		Logarithm of Rate Ratio	−0.02 (−0.71 - 0.65)

Appendix Table 5. Estimated difference in association of homelessness with hospital utilization and charges between study visits before the COVID-19 pandemic and after COVID-19 pandemic (March 1, 2020).

Setting	Outcome	Measure of Association	Interaction Estimate
Number of Admissions	Emergency Department	Rate Difference	0.23 (−1.03 - 2.6)
		Logarithm of Rate Ratio	0.26 (−0.33 - 0.84)
	Inpatient	Rate Difference	−0.01 (−0.16 - 0.15)
		Logarithm of Rate Ratio	0.09 (−0.47 - 0.67)
Total Charges	Emergency Department	Rate Difference	−367.91 (−1738.23 - 1529.5)
		Logarithm of Rate Ratio	0.01 (−0.53 - 0.54)
	Inpatient	Rate Difference	−189.36 (−3564.58 - 3211.61)
		Logarithm of Rate Ratio	−0.06 (−0.67 - 0.55)

Appendix Table 6. Estimated difference in association of homelessness with hospital utilization and charges between study visits during the COVID-19 pandemic and after during COVID-19 pandemic (March 1, 2020).

Setting	Outcome	Measure of Association	Interaction Estimate
Number of Admissions	Emergency Department	Rate Difference	0.05 (−2.69 - 2.71)
		Logarithm of Rate Ratio	0 (−0.9 - 0.85)
	Inpatient	Rate Difference	0.05 (−0.15 - 0.24)
		Logarithm of Rate Ratio	0.22 (−0.46 - 0.93)
Total Charges	Emergency Department	Rate Difference	−300.93 (−3259.62 - 2065.24)
		Logarithm of Rate Ratio	−0.04 (−0.81 - 0.7)
	Inpatient	Rate Difference	−853.55 (−6394.83 - 4313.56)
		Logarithm of Rate Ratio	−0.04 (−0.79 - 0.75)

Appendix Table 7. Adjusted relative rate and 95% bootstrap confidence interval for association of homelessness with emergency department and inpatient hospital admissions and charges with specific conditions present on discharge.

Cause Present During Hospitalization	Outcome	Setting	Relative Rate
Alcohol Use Disorders	Number of Admissions	Emergency Department	2.68 (1.85 - 3.96)
		Inpatient	1.39 (0.86 - 2.19)
	Total Charges	Emergency Department	2.06 (1.32 - 3.18)
		Inpatient	1.22 (0.66 - 2.13)
Depression	Number of Admissions	Emergency Department	2.81 (2.09 - 3.76)
		Inpatient	1.42 (0.92 - 2.15)
	Total Charges	Emergency Department	2.84 (2.04 - 3.86)
		Inpatient	1.41 (0.87 - 2.27)
Substance Use Disorders (non-Alcohol)	Number of Admissions	Emergency Department	2.25 (1.71 - 2.87)
		Inpatient	1.6 (1.23 - 2.11)
	Total Charges	Emergency Department	2.12 (1.66 - 2.62)
		Inpatient	1.55 (1.16 - 2.09)
Overdose	Number of Admissions	Emergency Department	2.28 (1.65 - 3.15)
		Inpatient	1.53 (0.78 - 2.91)
	Total Charges	Emergency Department	2.49 (1.8 - 3.48)
		Inpatient	1.85 (0.8 - 4.13)
Psychoses	Number of Admissions	Emergency Department	2.38 (1.14 - 3.86)
		Inpatient	2.54 (1.16 - 6.11)
	Total Charges	Emergency Department	2.1 (0.94 - 3.34)
		Inpatient	4.09 (1.72 - 9.18)

Appendix Table 8. Association of past-six-month homelessness with past-six-month emergency department and inpatient hospital utilization, by route of drug administration and pandemic era, excluding outliers^* (sensitivity analysis 1).

Outcome	Setting	Model	Rate Difference	Rate Ratio
Number of Admissions	Emergency Department	Crude	1.29 (1.03 - 1.56)	2.93 (2.52 - 3.38)
		Adjusted	0.97 (0.74 - 1.2)	1.98 (1.74 - 2.25)
	Inpatient	Crude	0.16 (0.1 - 0.22)	1.99 (1.61 - 2.44)
		Adjusted	0.11 (0.05 - 0.17)	1.5 (1.22 - 1.83)
Total Charges	Emergency Department	Crude	$1,705 ($1,297 - $2,139)	2.56 (2.15 - 3.02)
		Adjusted	$1,288 ($911 - $1,704)	1.86 (1.58 - 2.17)
	Inpatient	Crude	$2,708 ($1,347 - $4,212)	1.76 (1.37 - 2.23)
		Adjusted	$2,103 ($701 - $3,607)	1.5 (1.16 - 1.94)

^*Outliers were defined as 6-month periods when participants had more than 20 ED admissions or more than 5 IP admissions

Appendix Table 9. Past-six-month emergency department and hospital utilization and charges by housing status, route of drug administration, and pandemic-era among people who have injected drugs – sensitivity analysis 2 using 6-month period after study visit.

Outcome	Setting	Mean
		All Participants	No Homelessness	Homeless
Number of Admissions	Emergency Department	0.89	0.71	2.53
	Inpatient	0.19	0.18	0.31
Total Charges	Emergency Department	1402.99	1186.76	3441.59
	Inpatient	4626.33	4434.20	6437.64

Appendix Table 10. Association of past-six-month homelessness with past-six-month emergency department and inpatient hospital utilization, by route of drug administration and pandemic era – sensitivity analysis 2 using 6-month period after study visit.

Outcome	Setting	Model	Rate Difference	Rate Ratio
Number of Admissions	Emergency Department	Crude	1.76 (0.99 - 3.2)	3.45 (2.42 - 5.2)
		Adjusted	1.18 (0.57 - 2.28)	1.92 (1.49 - 2.42)
	Inpatient	Crude	0.14 (0.07 - 0.21)	1.76 (1.38 - 2.21)
		Adjusted	0.04 (−0.03 - 0.11)	1.15 (0.9 - 1.46)
Total Charges	Emergency Department	Crude	$2,180 ($1,345 - $3,587)	2.84 (2.14 - 3.93)
		Adjusted	$1,415 ($725 - $2,436)	1.73 (1.38 - 2.11)
	Inpatient	Crude	$1,957 ($143 - $4,117)	1.44 (1.03 - 1.96)
		Adjusted	$561 (−$1,356 - $2,863)	1.1 (0.79 - 1.54)

Appendix Figure 1. Histogram of distribution of past-six-month emergency department and hospital admissions and charges.